genienlp/decanlp/multiprocess/distributed_data_parallel.py

#
# Copyright (c) 2018, Salesforce, Inc.
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# * Redistributions of source code must retain the above copyright notice, this
#   list of conditions and the following disclaimer.
#
# * Redistributions in binary form must reproduce the above copyright notice,
#   this list of conditions and the following disclaimer in the documentation
#   and/or other materials provided with the distribution.
#
# * Neither the name of the copyright holder nor the names of its
#   contributors may be used to endorse or promote products derived from
#   this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

import torch
from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
import torch.distributed as dist
from torch.nn.modules import Module
import logging

logger = logging.getLogger(__name__)

class DistributedDataParallel(Module):

    def __init__(self, module):
        super(DistributedDataParallel, self).__init__()
        self.warn_on_half = True#$ True if dist._backend == dist.dist_backend.GLOO else False

        self.module = module

        for p in self.module.state_dict().values():
            if torch.is_tensor(p):
                dist.broadcast(p, 0)

        def allreduce_params():
            if(self.needs_reduction):
                self.needs_reduction = False
                buckets = {}
                for param in self.module.parameters():
                    if param.requires_grad and param.grad is not None:
                        tp = type(param.data)
                        if tp not in buckets:
                            buckets[tp] = []
                        buckets[tp].append(param)
                if self.warn_on_half:
                    if torch.cuda.HalfTensor in buckets:
                        logger.warning("gloo dist backend for half parameters may be extremely slow." +
                                       " It is recommended to use the NCCL backend in this case.")
                        self.warn_on_half = False

                for tp in buckets:
                    bucket = buckets[tp]
                    grads = [param.grad.data for param in bucket]
                    coalesced = _flatten_dense_tensors(grads)
                    dist.all_reduce(coalesced)
                    coalesced /= dist.get_world_size()
                    for buf, synced in zip(grads, _unflatten_dense_tensors(coalesced, grads)):
                        buf.copy_(synced)

        for param in list(self.module.parameters()):
            if param.requires_grad:
               def allreduce_hook(*unused):
                   param._execution_engine.queue_callback(allreduce_params)
               param.register_hook(allreduce_hook)

    def forward(self, *inputs, **kwargs):
        self.needs_reduction = True
        return self.module(*inputs, **kwargs)